Confidence-Aware Autonomous Knowledge Graph Evolution for Lifelong Robotic TAMP Systems

Neuro-symbolic Task and Motion Planning (TAMP) architectures anchor Large Language Model (LLM) reasoning within a Knowledge Graph (KG) K to eliminate plan-level hallucinations, but letting K evolve autonomously across missions risks contaminating it with confidently-wrong facts that both the planner and the perception stack will then trust. We formalize this problem and address it with a unified Autonomous Knowledge Update framework combining confidence-aware entity-attribute belief fusion, geolocated obstacle ingestion, and density-based emergence of Zones of Interest, all governed by the same hybrid syntactic-semantic Dempster-Shafer fusion rule so that contamination resistance is a property of the framework ratherthan of any single channel. We give a full algorithmic specification of its three update channels and a field evaluation over 8-mission made in a EU H2020 OLGA project context deployment that shows obstacle-driven mid-mission replanning triggers falling from 10 to 0 events, KG growth correlating with plan richness at r = 0.993, our fusion mechanism keeping contamination roughly an order of magnitude lower than a naive last-write-wins updater across tested observation error rates, and cross-checking the on-board recognition module against K raising its top-1 accuracy from 89.7% to 94.0% over the campaign.